Pith. sign in

REVIEW 3 major objections 5 minor 48 references

Cloud can drive only after uplink, then memory-bound VLA latency, then cost each clear in sequence.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-13 00:44 UTC pith:HFJEJCXA

load-bearing objection Solid systems map of when cloud AV inference is feasible; the year-by-year VLA wall is stack-specific, but the nested-regime framing and S2 economics still hold. the 3 major comments →

arxiv 2607.09045 v1 pith:HFJEJCXA submitted 2026-07-10 eess.SY cs.SY

Can the Cloud Drive? Infrastructure Feasibility of Offloading Autonomous Driving Across 5G and 6G

Pouya Parsa , Kawon Han , Seongjin Choi This is my paper
classification eess.SY cs.SY
keywords autonomous drivingvehicular edge computingtask offloading5G/6G communicationsvision-language-action modelsinfrastructure feasibilityroofline GPU modelutilization-aware cost
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Frontier autonomous-driving models, especially vision-language-action systems that cost tens of TFLOPs per decision, make full onboard hardware wasteful: peak chips sit idle most of the day. Cloud inference can share GPUs across active vehicles, but only if the vehicle can upload through a capacity-limited cell, reach a free GPU, and return a decision inside a closed-loop deadline. The paper couples communication limits, a roofline GPU service model, stochastic latency, and utilization-aware cost across three model classes, three offloading splits, and three radio generations, then applies the whole stack to New York City under a tight 100 ms reactive budget and a looser 300 ms deliberative tier. Three nested constraints bind in order: dense cells first make the uplink the bottleneck (5G fails early; 5G-Advanced is the practical threshold for feature-level offloading); under the reactive budget, near-term VLA remains latency-infeasible regardless of bandwidth because autoregressive decoding is memory-bandwidth-bound; only after both gates clear does utilization-pooled cloud cost undercut expensive idle onboard VLA hardware. Latency therefore decides which model is admissible in which year; cost decides whether it is economical.

Core claim

In a single dense-city setting the feasibility of cloud-driven autonomy is governed by three nested binding regimes. Communication binds first: 5G cannot sustain feature-level offloading under realistic loading, 5G-Advanced is the practical threshold, and 6G supplies headroom. Compute binds next under the 100 ms reactive budget: near-term VLA is latency-infeasible regardless of bandwidth because autoregressive FP16 decode is memory-bandwidth-bound (about 114 ms of a 153 ms floor on 2025 hardware); the floor clears 100 ms around 2027, after which 6G admits feature-level VLA by about 2028 while 5G-Advanced does so only at light loading. Cost binds last: once admissible, shared cloud GPUs under

What carries the argument

Three nested binding regimes produced by a joint analytical pipeline: interference-aware uplink admission, a roofline GPU service model that separates compute-bound encoder/prefill from HBM-bandwidth-bound autoregressive decode, stochastic tail-latency feasibility under 100 ms reactive and 300 ms deliberative budgets, and utilization-aware total-cost-of-ownership crossover.

Load-bearing premise

The claim that near-term VLA stays latency-infeasible under a 100 ms reactive loop rests on the premise that service time remains dominated by fully autoregressive FP16 decode whose memory-bandwidth floor is accurately given by the paper's roofline calibration.

What would settle it

Measure end-to-end reactive-loop latency of a production VLA stack on 2025-class hardware that uses FP8 weights or a parallel (diffusion/flow) action decoder; if the decode floor falls well below 100 ms before 2027, the compute-bound regime and all year-by-year admissibility claims collapse.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper asks whether cloud/edge inference can economically drive autonomous vehicles under closed-loop latency constraints. It couples three systems—an offloading-strategy spectrum (S1 raw-sensor / S2 feature-level / S3 query-level) for E2E, VLM, and VLA models; an interference-aware 5G/5G-Advanced/6G uplink admission model; and a roofline GPU service model with stochastic tail latency and utilization-aware TCO—into a joint feasibility gate and cost-crossover analysis applied to a 1,296-branch New York City matrix. Under a reactive 100 ms budget and a deliberative 300 ms tier (the latter only behind an onboard reactive fallback), the authors report three nested binding regimes: communication binds first in dense cells (5G fails early; 5G-Advanced is the practical threshold for S2), compute binds next for near-term VLA because autoregressive FP16 decode is memory-bandwidth-bound (~114 ms on 2025 hardware; floor clears ~2027), and cost binds last once a branch is admissible (utilization-pooled cloud GPUs undercut expensive idle VLA onboard hardware, with the crossover concentrating at S2). Latency decides which model is admissible in which year; cost decides whether it is economical.

Significance. If the nested-regime picture holds, the paper supplies a concrete, operator-facing deployment sequence that the three literatures (driving models, vehicular communications, vehicular edge computing) have not previously integrated. The framework is internally consistent, parameterizes independently sourced inputs (3GPP/ITU link budgets, NVIDIA datasheets, DOE parking statistics, published model FLOPs), and produces falsifiable year/generation admissibility maps and cost crossovers rather than a binary feasibility claim. The explicit separation of reactive and deliberative budgets, the residual-TOPS accounting for S1–S3, and the utilization-driven TCO comparison are useful contributions for spectrum planning, shared-edge co-investment, and safety certification of cloud-dependent AV stacks. The main result is conditional on a conservative decoder stack, but the paper itself flags that condition and the framework can absorb alternative architectures by updating Table 2 and Eq. (9).

major comments (3)
  1. §4.2 and Eq. (9): The compute-bound regime—and all subsequent year/generation VLA admissibility numbers (0/432 reactive branches in 2026; floor clears ~2027; 6G admits VLA-S2 ~2028; 5G-Advanced only at light loading)—rests on the premise that service time remains dominated by fully autoregressive FP16 decode (~114 ms HBM-bound on B300). The paper correctly notes that FP8 roughly halves the per-token read and that parallel diffusion/flow action decoders remove the autoregressive re-reads, moving reactive admissibility years earlier. Because this is load-bearing for the nesting claim, the main results need at least a one- or two-scenario sensitivity (e.g., FP8 and a non-AR decoder) so readers can see which regime boundaries survive when the conservative stack is relaxed. Without that, the year-by-year map is presented as more structural than the Limitations section itself allows.
  2. §3.8 / Eq. (28) and §4.3 / Fig. 10: Deliberative-tier cost ratios charge only each strategy’s residual onboard hardware and explicitly exclude the onboard reactive-fallback controller that the 300 ms tier requires by construction (§3.5.3, §5). The manuscript labels the resulting VLA cost-attractive region an “optimistic bound,” but the bound is not quantified. Because the paper’s economic takeaway is that S2 concentrates the VLA crossover, a simple additive fallback cost (or a sensitivity band) is needed so the deliberative TCO comparison is not systematically biased toward cloud.
  3. §3.4.3 / Table 4 and Eqs. (10)–(11): GPU and HBM evolution are calibrated to four A100→B300 points with a decelerating-rate model (r0_ϕ=64%/yr → r∞_ϕ=10%/yr, λ=0.15). The 2027 floor-clearing year and the 2028 6G-admission claim inherit this extrapolation. A short sensitivity on the long-run floors (or on λ) would show how much the compute-regime timeline moves under slower or faster memory-bandwidth growth; without it the year labels are more precise than the calibration supports.
minor comments (5)
  1. Table 3 / residual H_s: The residual INT8 TOPS for VLA under S2 (550) and S3 (2900) versus the 3000 TOPS full baseline should be cross-checked against the phase FLOPs in Table 2 and the 2:1 INT8/FP16 conversion; a one-sentence derivation would help reproducibility.
  2. Fig. 3 and Fig. 5: Axis labels and the capacity-cliff annotation are dense; a short caption note defining N_max_c and Δ would improve readability for non-communications readers.
  3. §3.5.2: The processor-sharing approximation for L_nq and the Chernoff/MGF tail for Eq. (19) are standard but cited lightly; a pointer to the exact MGF forms used would aid replication.
  4. References: Alpamayo is cited as a CES 2026 / Hugging Face release; ensure the archival citation is stable before camera-ready. A few 6G V2X surveys already in the bibliography could be tied more explicitly to the uplink-heavy framing in §5.2.
  5. Notation: μ_eff(s,m,t) is used both as service rate and (inverted) as service time in Eq. (9); a single convention would reduce friction.

Circularity Check

1 steps flagged

No load-bearing circularity: nested regimes are computed from external standards, datasheets, and published model FLOPs; GPU-evolution rates are fitted historical inputs used transparently as scenario parameters, not first-principles predictions.

specific steps
  1. fitted input called prediction [Sec. 3.4.3 Eqs. (10)–(11), Table 4; Sec. 4.2 / abstract claim that VLA floor clears 100 ms around 2027]
    "we let GPU performance follow the historical A100→H100→B300 trend and project forward from the B300 baseline year t0 = 2025... Table 4 calibrates these rates against NVIDIA A100–B300 2020–2025 CAGRs: ∼64%/yr compute, ∼31%/yr memory bandwidth. The fitted deceleration model sets r0_ϕ = 64%/yr and r0_B = 31%/yr... Its floor clears 100 ms around 2027"

    Growth rates are fitted to historical GPU datasheet CAGRs, then accumulated to project future HBM bandwidth and thus the year the VLA deterministic floor drops below 100 ms. The ~2027 date is therefore forced by the fitted trajectory rather than independently measured or derived from first principles. Mild only: the paper treats the rates as scenario inputs and does not claim to predict GPU evolution itself; feasibility years are conditional projections, not circular re-statements of the fit target.

full rationale

The paper's central claims (three nested binding regimes; year/generation admissibility of VLA; S2 cost crossover) are outputs of an analytical pipeline whose inputs are independently sourced: 3GPP/ITU link budgets and spectral-efficiency models, NVIDIA A100–B300 datasheets, DOE parking/utilization statistics, and published E2E/VLM/VLA FLOPs (UniAD, DriveLM, Alpamayo). Communication capacity (Eqs. 1–5), the roofline service time (Eq. 9), stochastic tail bounds (Eqs. 14–19), the joint feasibility indicator (Eq. 21), facility-location provisioning (Eqs. 22–27), and the TCO crossover (Eq. 28) are not defined in terms of the regimes they produce; the regimes are the result of applying those gates to the NYC scenario grid. The only mild fitted-input step is the decelerating GPU-evolution model (Sec. 3.4.3, Eqs. 10–11, Table 4), calibrated to historical CAGRs and then used to project when the VLA deterministic floor clears 100 ms (~2027). That projection is conditional on the fitted rates and is presented as scenario analysis, not as a first-principles prediction of hardware or of feasibility independent of those rates; the paper itself flags that FP8 or non-autoregressive decoders would move the wall years earlier. There is no self-definitional loop, no uniqueness theorem imported from the authors, no ansatz smuggled via self-citation, and no renaming of a known empirical pattern as a derived law. Score 1 reflects only the transparent historical fit used for forward projection; the derivation chain is otherwise self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 2 invented entities

The central nested-regime claim rests on a large but mostly standard set of domain parameters (3GPP/ITU link budgets, NVIDIA GPU trends, published model FLOPs, DOE utilization) plus a handful of modeling choices (100/300 ms budgets, S1/S2/S3 split points, decelerating CAGR form, static utilization). No new physical entities are postulated; free parameters are scenario inputs rather than fitted to force the conclusion. The ledger is therefore heavy on domain assumptions and free parameters but light on invented entities.

free parameters (6)
  • GPU compute CAGR initial/asymptotic (r0_phi, r_inf_phi) = 64 %/yr → 10 %/yr, λ=0.15
    Fitted to A100–B300 history (64 %/yr → 10 %/yr) with deceleration λ=0.15; directly sets when VLA floor clears 100 ms.
  • HBM bandwidth CAGR (r0_B, r_inf_B) = 31 %/yr → 7 %/yr
    Same calibration (31 %/yr → 7 %/yr); dominates the VLA decode term that creates the compute-bound regime.
  • VLA baseline onboard cost = $8 500/vehicle-year (2026)
    Sets the $8 500/vehicle-year figure against which cloud crossover is measured; declines 15 %/yr by assumption.
  • Strategy uplink rates B_s = 100 / 25 / 3 Mbps
    Chosen representative values (100/25/3 Mbps) inside published ranges; control the communication gate.
  • Reactive / deliberative latency budgets = 100 ms / 300 ms
    100 ms from 10 Hz planning; 300 ms chosen as deliberative tier; the 100 ms value is the binding parameter for the compute regime.
  • Tail probability ε and queue over-provisioning = ε=10^{-5}
    ε=10^{-5} and resulting 15–25 % GPU over-provisioning are design choices that affect feasibility counts.
axioms (6)
  • domain assumption Interference-aware OFDMA uplink SINR and spectral efficiency follow the residual-interference form of Andrews et al. / 3GPP UMa (Eqs. 2–4).
    Standard cellular model; parameters taken from 3GPP TR 38.901 and related literature.
  • domain assumption Autoregressive VLA service time decomposes into compute-bound encoder/prefill plus HBM-bound decode (roofline Eq. 9).
    Standard roofline applied to LLM-style decoding; calibrated to one stack (Alpamayo / TensorRT-LLM).
  • standard math GPU queueing is M/D/c with Erlang-C tail bound (Eq. 8).
    Classical multi-server queueing; used for provisioning.
  • domain assumption 6G peak rates, latency, and density equal ITU IMT-2030 aspirational targets.
    Explicitly flagged as aspirational in limitations; if realized capacity is lower the urban frontier moves.
  • domain assumption Personal-vehicle utilization baseline u=0.05 from DOE parking statistics; higher u values represent mixed/robotaxi fleets.
    External empirical input that drives the cost-crossover utilization effect.
  • ad hoc to paper Deliberative 300 ms tier is admissible only behind an onboard reactive fallback that closes the 100 ms loop locally.
    Modeling choice that reclassifies cloud as non-safety-critical planning aid; required for the dual-budget analysis.
invented entities (2)
  • Three nested binding regimes (communication → compute → cost) no independent evidence
    purpose: Organizing frame that sequences the feasibility gates and explains why S2 is the robust middle point.
    Conceptual taxonomy introduced by the paper; not a physical entity, but the central interpretive claim. Independent evidence is the quantitative branch counts and year-by-year floors derived from the model.
  • S1/S2/S3 offloading spectrum with residual TOPS and cloud FLOPs tables independent evidence
    purpose: Concrete split points that map AD pipeline stages to uplink demand and leftover onboard hardware.
    Standard idea of split computing, instantiated with specific numbers from UniAD/DriveLM/Alpamayo; the particular residual-TOPS values are paper-specific inputs.

pith-pipeline@v1.1.0-grok45 · 33714 in / 4068 out tokens · 44957 ms · 2026-07-13T00:44:20.799122+00:00 · methodology

0 comments
read the original abstract

Frontier autonomous-driving models -- especially vision-language-action (VLA) models, whose forward pass approaches $\sim$60~TFLOPs -- are outgrowing economical onboard deployment, since peak hardware sits idle most of the day. Cloud inference can instead share GPUs across active vehicles, but the vehicle must upload through a capacity-limited uplink, reach a GPU without queueing, and return a decision within the closed-loop budget. This paper asks: can the cloud drive? We answer with an analytical framework coupling communication limits, a roofline GPU service model, stochastic latency, and utilization-aware cost across three model classes, three offloading strategies, and three communication generations, applied to New York City. Separating a reactive 100~ms budget from a 300~ms deliberative tier (presuming an onboard reactive fallback), we find three \emph{nested} binding regimes. Communication binds first in dense cells: 5G fails early, 5G-Advanced is the practical threshold for feature-level offloading, and 6G adds headroom. Compute binds next under the reactive budget: near-term VLA is latency-infeasible regardless of bandwidth, because autoregressive FP16 decode is memory-bandwidth-bound (~114 ms on 2025 hardware). Its floor clears 100 ms around 2027; 6G then admits feature-level VLA by ~2028, 5G-Advanced only at light loading and not the dense corridor, and the deliberative tier from 2026. Cost binds last: once admissible, utilization-pooled cloud GPUs undercut onboard hardware for VLA, whose baseline (up to \$8,500 per vehicle-year) is expensive and idle; feature-level offloading (S2) is where the VLA cost crossover concentrates. Latency decides which model is admissible in which year; cost decides whether it is economical.

Figures

Figures reproduced from arXiv: 2607.09045 by Kawon Han, Pouya Parsa, SeongJin Choi.

Figure 1
Figure 1. Figure 1: Analytical pipeline. The Communication Requirements Model defines the communication-side limit, and the GPU Service Model defines the cloud-side service model used by the Stochastic Latency Model. Those results then feed the Feasibility Threshold; feasible branches proceed to edge infrastructure optimization and are finally compared through the Total Cost of Ownership and Crossover step. and the cloud GPU … view at source ↗
Figure 2
Figure 2. Figure 2: Offloading strategy spectrum. 3.2.4. In-Vehicle Baseline We denote the peak full-pipeline onboard requirements for model class 𝑚 by 𝐻full 𝑚 , allowing later cost equations to price the hardware each vehicle must carry for full in-vehicle inference. For this baseline, full-pipeline onboard requirements are 𝐻full E2E = 85 TOPS, 𝐻full VLM = 1235 TOPS, and 𝐻full VLA = 3,000 TOPS. This notation defines the full… view at source ↗
Figure 3
Figure 3. Figure 3: Bandwidth-capacity tradeoff. Colored lines show the required uplink demand for S1–S3 as vehicles per cell increase, while gray generation curves show the available uplink capacity under the interference-aware cell model. Intersections mark the maximum bandwidth-feasible vehicles per cell. Formally, the required GPU count is defined as the smallest 𝑐𝑒 satisfying the queueing-tail constraint, 𝑐 ∗ 𝑒 (𝑠, 𝑚, 𝑡)… view at source ↗
Figure 4
Figure 4. Figure 4: End-to-end control loop delay decomposition. Solid borders denote deterministic components; dashed borders denote stochastic components that consume the remaining latency slack. 3.5.1. Deterministic Latency Floor For latency decomposition, sensing delay (𝐿s ) isolates the fixed local latency that precedes vehicle-to-network communication and cloud inference. We model this fixed sensing delay as a 5 ms term… view at source ↗
Figure 5
Figure 5. Figure 5: Tail latency under increasing cell utilization. Curves show the 99.999th-percentile end-to-end loop delay under 5G-Advanced (compact E2E workload) as vehicles per cell increase for S1–S3. Horizontal lines mark the reactive 100 ms and deliberative 300 ms budgets. The sharp rise at high loading is the capacity cliff discussed in the text: stochastic network scheduling consumes the remaining budget Δ, so poin… view at source ↗
Figure 6
Figure 6. Figure 6: Case Study 1: NYC communication frontier. The y-axis gives the maximum citywide fleet penetration that remains bandwidth-feasible at each utilization. Reference points: 1% = 22K vehicles, 10% = 220K vehicles. the jointly feasible set to 487 branches in 2026; under the 300 ms deliberative budget, 903 branches are jointly feasible. The gap between these two counts is the compute-bound regime made quantitativ… view at source ↗
Figure 7
Figure 7. Figure 7: Case Study 1: NYC communication sweep. Each cell shows the minimum communication generation required to keep a strategy bandwidth-feasible for the given AV penetration and utilization pair. The highlighted cell corresponds to the reference point of 10% penetration and 𝑢 = 0.45. fit within the 6G envelope, and many moderate combinations are already reachable with 5G-Advanced. Because the figure screens comm… view at source ↗
Figure 8
Figure 8. Figure 8: Regime 2 (compute-bound). (a) Deterministic latency floor 𝐿det versus GPU year for VLA (S1–S3), with VLM-S2 and E2E-S2 references; the reactive 100 ms and deliberative 300 ms budgets are marked. The VLA floor first drops below 100 ms in 2027, but this is a lower bound: once the access-scheduling tail is added, reactive admissibility is later and generation-dependent (Section 4.2); the floor is under 300 ms… view at source ↗
Figure 9
Figure 9. Figure 9: Regime 3 (cost-bound), reactive budget. NYC per-vehicle hybrid cost versus the in-vehicle baseline (black) over time at 5G-Advanced. Rows: model classes. Columns: utilization. Solid curves are latency-feasible under the 100 ms budget; dashed segments are latency-infeasible. VLA cloud (S2) falls below the $8,500 baseline once it clears the compute regime. supplies so much headroom that raw-sensor offloading… view at source ↗
Figure 10
Figure 10. Figure 10: Regime 3 (cost-bound), deliberative tier in 2028, where VLA is latency-admissible. Cheapest feasible NYC cloud cost divided by the in-vehicle baseline; cell text gives the cheapest feasible strategy and gray cells are infeasible. Green marks cloud-cheaper cells. For VLA the cost-attractive region is large and dominated by S2. strategy there shifts to S3, which keeps more compute onboard and rarely beats t… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

48 extracted references · 15 canonical work pages · 2 internal anchors

  1. [1]

    Y. Hu, J. Yang, L. Chen, K. Li, C. Sima, X. Zhu, S. Chai, S. Du, T. Lin, W. Wang, L. Lu, X. Jia, Q. Liu, J. Dai, Y. Qiao, H. Li, Planning- oriented autonomous driving, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 17853–17862.doi:10.1109/CVPR52729.2023.01712

    doi:10.1109/cvpr52729.2023.01712 2023

  2. [2]

    Jiang, S

    B. Jiang, S. Chen, Q. Xu, B. Liao, J. Chen, H. Zhou, Q. Zhang, W. Liu, C. Huang, X. Wang, VAD: Vectorized scene representation for efficientautonomousdriving,in:ProceedingsoftheIEEE/CVFInternationalConferenceonComputerVision(ICCV),2023,pp.8306–8316. doi:10.1109/ICCV51070.2023.00766

    doi:10.1109/iccv51070.2023.00766 2023

  3. [3]

    C. Sima, K. Renz, K. Chitta, L. Chen, H. Zhang, C. Xie, J. Beißwenger, P. Luo, A. Geiger, H. Li, DriveLM: Driving with graph visual question answering, in: Proceedings of the European Conference on Computer Vision (ECCV), 2024, pp. 256–274.doi:10.1007/ 978-3-031-72943-0_15

    2024

  4. [4]

    Hwang, R

    J.-J. Hwang, R. Xu, H. Lin, W.-C. Hung, J. Ji, K. Choi, D. Huang, T. He, P. Covington, B. Sapp, Y. Zhou, J. Guo, D. Anguelov, M. Tan, EMMA: End-to-end multimodal model for autonomous driving, Transactions on Machine Learning Research (2025)

    2025

  5. [5]

    Model available athttps://huggingface.co/nvidia/Alpamayo-R1-10B(Jan

    NVIDIA Corporation, Alpamayo: Open-source vision-language-action models for autonomous vehicle development, NVIDIA Newsroom, announced at CES 2026. Model available athttps://huggingface.co/nvidia/Alpamayo-R1-10B(Jan. 2026). URLhttps://nvidianews.nvidia.com/news/alpamayo-autonomous-vehicle-development

    2026

  6. [6]

    U.S. DOE, Fact of the Week #1356: Household Vehicles Were Parked 95% of a Typical Day in 2022,https://www.energy.gov/eere/ vehicles/articles/fotw-1356-august-19-2024-household-vehicles-were-parked-95-typical-day-2022,basedon2022 National Household Travel Survey data (Aug. 2024)

    2022

  7. [7]

    L.Chen,P.Wu,K.Chitta,B.Jaeger,A.Geiger,H.Li,End-to-endautonomousdriving:Challengesandfrontiers,IEEETransactionsonPattern Analysis and Machine Intelligence 46 (12) (2024) 10164–10183.doi:10.1109/TPAMI.2024.3435937

    doi:10.1109/tpami.2024.3435937 2024

  8. [8]

    S. Liu, L. Liu, J. Tang, B. Yu, Y. Wang, W. Shi, Edge computing for autonomous driving: Opportunities and challenges, Proceedings of the IEEE 107 (8) (2019) 1697–1716.doi:10.1109/JPROC.2019.2915983

    doi:10.1109/jproc.2019.2915983 2019

  9. [9]

    G. Naik, B. Choudhury, J.-M. Park, IEEE 802.11bd & 5G NR V2X: Evolution of radio access technologies for V2X communications, IEEE Access 7 (2019) 70169–70184.doi:10.1109/ACCESS.2019.2919489

    doi:10.1109/access.2019.2919489 2019

  10. [10]

    M.H.C.Garcia,A.Molina-Galan,M.Boban,J.Gozalvez,B.Coll-Perales,T.Sahin,A.Kousaridas,Atutorialon5GNRV2Xcommunications, IEEE Communications Surveys & Tutorials 23 (3) (2021) 1972–2026.doi:10.1109/COMST.2021.3057017

    doi:10.1109/comst.2021.3057017 2021

  11. [11]

    Lin, An overview of 5G advanced evolution in 3GPP release 18, IEEE Communications Standards Magazine 6 (3) (2022) 77–83

    X. Lin, An overview of 5G advanced evolution in 3GPP release 18, IEEE Communications Standards Magazine 6 (3) (2022) 77–83. doi:10.1109/MCOMSTD.0001.2200001

    work page doi:10.1109/mcomstd.0001.2200001 2022

  12. [12]

    W. Chen, X. Lin, J. Lee, A. Toskala, S. Sun, C. F. Chiasserini, L. Liu, 5G-advanced towards 6G: Past, present, and future, arXiv preprint arXiv:2303.07456 (2023).doi:10.48550/arXiv.2303.07456

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2303.07456 2023

  13. [13]

    Noor-A-Rahim, Z

    M. Noor-A-Rahim, Z. Liu, H. Lee, M. O. Khyam, J. He, D. Pesch, K. Moessner, W. Saad, H. V. Poor, 6G for vehicle-to-everything (V2X) communications: Enabling technologies, challenges, and opportunities, Proceedings of the IEEE 110 (6) (2022) 712–734.doi: 10.1109/JPROC.2022.3173031

    work page doi:10.1109/jproc.2022.3173031 2022

  14. [14]

    A. B. de Souza, P. A. L. Rego, T. Carneiro, J. d. C. Rodrigues, P. P. Reboucas Filho, J. N. de Souza, V. Chamola, V. H. C. de Albuquerque, B. Sikdar, Computation offloading for vehicular environments: A survey, IEEE Access 8 (2020) 198214–198243.doi:10.1109/ACCESS. 2020.3033828

    doi:10.1109/access 2020

  15. [15]

    Ahmed, S

    M. Ahmed, S. Raza, M. A. Mirza, A. Aziz, M. A. Khan, W. U. Khan, J. Li, Z. Han, A survey on vehicular task offloading: Classification, issues, and challenges, Journal of King Saud University - Computer and Information Sciences 34 (7) (2022) 4135–4162.doi:10.1016/j. jksuci.2022.05.016

    doi:10.1016/j 2022

  16. [16]

    Karimi, Y

    E. Karimi, Y. Chen, B. Akbari, Task offloading in vehicular edge computing networks via deep reinforcement learning, Computer Communications 189 (2022) 193–204.doi:10.1016/j.comcom.2022.04.006

    work page doi:10.1016/j.comcom.2022.04.006 2022

  17. [17]

    15449–15458.doi:10.1109/CVPR52733

    X.Weng,B.Ivanovic,Y.Wang,Y.Wang,M.Pavone,PARA-Drive:Parallelizedarchitectureforreal-timeautonomousdriving,in:Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 15449–15458.doi:10.1109/CVPR52733. 2024.01463

    doi:10.1109/cvpr52733 2024

  18. [18]

    K.Chitta,A.Prakash,B.Jaeger,Z.Yu,K.Renz,A.Geiger,Transfuser:Imitationwithtransformer-basedsensorfusionforautonomousdriving, IEEE Transactions on Pattern Analysis and Machine Intelligence 45 (11) (2023) 12878–12895.doi:10.1109/TPAMI.2022.3200245

    doi:10.1109/tpami.2022.3200245 2023

  19. [19]

    X. Tian, J. Gu, B. Li, Y. Liu, Y. Wang, Z. Zhao, K. Zhan, P. Jia, X. Lang, H. Zhao, DriveVLM: The convergence of autonomous driving and large vision-language models, in: Proceedings of the Conference on Robot Learning (CoRL), 2024

    2024

  20. [20]

    H. Shao, Y. Hu, L. Wang, S. L. Waslander, Y. Liu, G. Song, H. Li, LMDrive: Closed-loop end-to-end driving with large language models, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 15120–15130. doi:10.1109/CVPR52733.2024.01432

    doi:10.1109/cvpr52733.2024.01432 2024

  21. [21]

    Parsa et al.:Preprint submitted to ElsevierPage 24 of 25 Can the Cloud Drive?

    Z.Xu,Y.Zhang,E.Xie,Z.Zhao,Y.Guo,K.-Y.K.Wong,Z.Li,H.Zhao,DriveGPT4:Interpretableend-to-endautonomousdrivingvialarge language model, IEEE Robotics and Automation Letters 9 (10) (2024) 8186–8193.doi:10.1109/LRA.2024.3440097. Parsa et al.:Preprint submitted to ElsevierPage 24 of 25 Can the Cloud Drive?

    doi:10.1109/lra.2024.3440097 2024

  22. [22]

    Jiang, D

    S.Jiang,Z.Huang,K.Qian,Z.Luo,T.Zhu,Y.Zhong,Y.Tang,M.Kong,Y.Wang,S.Jiao,H.Ye,Z.Sheng,X.Zhao,T.Wen,Z.Fu,S.Chen, K. Jiang, D. Yang, S. Choi, L. Sun, A survey on vision-language-action models for autonomous driving, in: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 4583–4595.doi:10.1109/ICCVW69036.2025.00476

    doi:10.1109/iccvw69036.2025.00476 2025

  23. [23]

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. Sanketi, Q. Vuong, T. Kollar, B. Burchfiel, R. Tedrake, D. Sadigh, S. Levine, P. Liang, C. Finn, OpenVLA: An open-source vision-language-action model, in: Proceedings of the Conference on Robot Learning (CoRL), 2024

    2024

  24. [24]

    R.DhineshKumar,A.Rammohan,Revolutionizingintelligenttransportationsystemswithcellularvehicle-to-everything(C-V2X)technology: Currenttrends,usecases,emergingtechnologies,standardizationbodies,industryanalyticsandfuturedirections,VehicularCommunications 43 (2023) 100638.doi:10.1016/j.vehcom.2023.100638

    work page doi:10.1016/j.vehcom.2023.100638 2023

  25. [25]

    3GPP, NR: Overall description; stage-2, Tech. Rep. TS 38.300 V17.0.0, 3rd Generation Partnership Project (2022)

    2022

  26. [26]

    ITU-R, Framework and overall objectives of the future development of IMT for 2030 and beyond, Recommendation ITU-R M.2160-0, International Telecommunication Union Radiocommunication Sector (Nov. 2023). URLhttps://www.itu.int/rec/R-REC-M.2160-0-202311-I/en

    2030

  27. [27]

    W.Saad,M.Bennis,M.Chen,Avisionof6Gwirelesssystems:Applications,trends,technologies,andopenresearchproblems,IEEENetwork 34 (3) (2020) 134–142.doi:10.1109/MNET.001.1900287

    work page doi:10.1109/mnet.001.1900287 2020

  28. [28]

    S. Dang, O. Amin, B. Shihada, M.-S. Alouini, What should 6G be?, Nature Electronics 3 (1) (2020) 20–29.doi:10.1038/ s41928-019-0355-6

    2020

  29. [29]

    H.Tataria,M.Shafi,A.F.Molisch,M.Dohler,H.Sjöland,F.Tufvesson,6Gwirelesssystems:Vision,requirements,challenges,insights,and opportunities, Proceedings of the IEEE 109 (7) (2021) 1166–1199.doi:10.1109/JPROC.2021.3061701

    doi:10.1109/jproc.2021.3061701 2021

  30. [30]

    H. Guo, X. Zhou, J. Liu, Y. Zhang, Vehicular intelligence in 6G: Networking, communications, and computing, Vehicular Communications 33 (2022) 100399.doi:10.1016/j.vehcom.2021.100399

    work page doi:10.1016/j.vehcom.2021.100399 2022

  31. [31]

    J.Santa,K.V.Katsaros,L.Bernal-Escobedo,S.Zougari,M.Miranda,O.Castañeda,B.Dalet,A.Amditis,Evaluationplatformfor5Gvehicular communications, Vehicular Communications 38 (2022) 100537.doi:10.1016/j.vehcom.2022.100537

    work page doi:10.1016/j.vehcom.2022.100537 2022

  32. [32]

    Y. Mao, C. You, J. Zhang, K. Huang, K. B. Letaief, A survey on mobile edge computing: The communication perspective, IEEE Communications Surveys & Tutorials 19 (4) (2017) 2322–2358.doi:10.1109/COMST.2017.2745201

    doi:10.1109/comst.2017.2745201 2017

  33. [33]

    Y. Li, C. Yang, X. Chen, Y. Liu, Mobility and dependency-aware task offloading for intelligent assisted driving in vehicular edge computing networks, Vehicular Communications 45 (2024) 100720.doi:10.1016/j.vehcom.2023.100720

    work page doi:10.1016/j.vehcom.2023.100720 2024

  34. [34]

    X.Liu,J.Zheng,Y.Li,M.Zhang,R.Wang,Y.He,Multi-pathserialtasksoffloadingstrategyanddynamicschedulingoptimizationinvehicular edge computing networks, Vehicular Communications 49 (2024) 100827.doi:10.1016/j.vehcom.2024.100827

    doi:10.1016/j.vehcom.2024.100827 2024

  35. [35]

    S. Wang, D. He, M. Yang, L. Duo, Cost-aware task offloading in vehicular edge computing: A Stackelberg game approach, Vehicular Communications 49 (2024) 100807.doi:10.1016/j.vehcom.2024.100807

    work page doi:10.1016/j.vehcom.2024.100807 2024

  36. [36]

    K.Hejja,S.Berri,H.Labiod,Networkslicingwithload-balancingfortaskoffloadinginvehicularedgecomputing,VehicularCommunications 34 (2022) 100419.doi:10.1016/j.vehcom.2021.100419

    doi:10.1016/j.vehcom.2021.100419 2022

  37. [37]

    A.Mekrache,A.Bradai,E.Moulay,S.Dawaliby,Deepreinforcementlearningtechniquesforvehicularnetworks:Recentadvancesandfuture trends towards 6G, Vehicular Communications 33 (2022) 100398.doi:10.1016/j.vehcom.2021.100398

    work page doi:10.1016/j.vehcom.2021.100398 2022

  38. [38]

    Karolemeas, S

    C. Karolemeas, S. Tsigdinos, E. Moschou, K. Kepaptsoglou, Shared autonomous vehicles and agent based models: a review of methods and impacts, European Transport Research Review 16 (2024) 25.doi:10.1186/s12544-024-00644-2. URLhttps://doi.org/10.1186/s12544-024-00644-2

    work page doi:10.1186/s12544-024-00644-2 2024

  39. [39]

    Y. Yang, Z. Yuan, F. Xin, Y. Wang, D. Sun, Optimization model of taxi fleet size based on gps tracking data, Sustainability 11 (2019) 731. doi:10.3390/su11030731

    work page doi:10.3390/su11030731 2019

  40. [40]

    rep., NVIDIA Corporation, 600 W TDP, 96 GB GDDR7, 4,000 TOPS AI (FP4) (Mar

    NVIDIA Corporation, NVIDIA RTX PRO 6000 Blackwell Workstation Edition datasheet, Tech. rep., NVIDIA Corporation, 600 W TDP, 96 GB GDDR7, 4,000 TOPS AI (FP4) (Mar. 2025). URLhttps://www.nvidia.com/en-us/products/workstations/professional-desktop-gpus/rtx-pro-6000/

    2025

  41. [41]

    J.G.Andrews,F.Baccelli,R.K.Ganti,Atractableapproachtocoverageandrateincellularnetworks,IEEETransactionsonCommunications 59 (11) (2011) 3122–3134.doi:10.1109/TCOMM.2011.100411.100541

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1109/tcomm.2011.100411.100541 2011

  42. [42]

    3GPP, Study on channel model for frequencies from 0.5 to 100 GHz, Tech. Rep. TR 38.901 V16.1.0, 3rd Generation Partnership Project (2020)

    2020

  43. [43]

    Borst, N

    S. Borst, N. Hegde, Integration of streaming and elastic traffic in wireless networks, in: IEEE INFOCOM 2007, 2007, pp. 1884–1892. doi:10.1109/INFCOM.2007.219

    work page doi:10.1109/infcom.2007.219 2007

  44. [44]

    NHTSA,Automateddrivingsystems2.0:Avisionforsafety,Tech.rep.,U.S.DepartmentofTransportation,NationalHighwayTrafficSafety Administration (2017)

    2017

  45. [45]

    DOT, Ensuring American leadership in automated vehicle technologies: Automated vehicles 4.0, Tech

    U.S. DOT, Ensuring American leadership in automated vehicle technologies: Automated vehicles 4.0, Tech. rep., U.S. Department of Transportation (2020)

    2020

  46. [46]

    Leviathan, M

    Y. Leviathan, M. Kalman, Y. Matias, Fast inference from transformers via speculative decoding, in: International Conference on Machine Learning (ICML), PMLR, 2023, pp. 19274–19286

    2023

  47. [47]

    W.Fedus,B.Zoph,N.Shazeer,Switchtransformers:Scalingtotrillionparametermodelswithsimpleandefficientsparsity,JournalofMachine Learning Research 23 (120) (2022) 1–39

    2022

  48. [48]

    Hinton, O

    G. Hinton, O. Vinyals, J. Dean, Distilling the knowledge in a neural network, arXiv preprint arXiv:1503.02531 (2015). Parsa et al.:Preprint submitted to ElsevierPage 25 of 25

    Pith/arXiv arXiv 2015